| 
| 
| 
  
concept-level nodes each contained ten training 
instances with the exception of the Turf/Grass 
(TG) class which contained eleven instances and 
the Barren 1 (Bl) class with thirteen 
instances. The remaining 147 instances were 
then classified using SX-WEB's evaluation 
function. Of the 147 instances, 145 were 
classified correctly. One misclassification 
placed a Shrub Swamp (SS) instance into the 
Marsh (MA) class. The second misclassification 
placed a Southern Deciduous (SD) pixel into the 
Barren 1 (Bl) concept class. The overall 
classification accuracy was 98.6%. 
In the second experiment, we used sixty 
training instances, with four training 
instances being randomly chosen from each of 
the fifteen classes. Classification accuracy 
was 92.1%, with 223 of the 242 instances being 
correctly classified. When the 15 categories 
are generalized to seven categories [Urban 
(UR), Agricultural (AG), Deciduous (DE), 
Coniferous (CO), Water (WA), Wetland (WE), and 
Barren (BA)], so as to match the categories 
which were utilized in (Civco, 1992a), the 
accuracy rate increased to 96.2%, indicating 
that several of the misclassifications placed 
instances into similar concept categories. 
In the third experiment the number of training 
pixels was reduced to 45 (3 randomly-selected 
pixels for each of the fifteen categories), and 
257 pixels were then classified. Even with this 
small training set, 220 (85.6%) of the 257 
pixels in the testing set were correctly 
classified. The specific incorrect  clas- 
sifications can be identified in the confusion 
matrix of Figure 4. 
When the 15 categories are generalized to the 
seven categories of  (Civco, 1992a), the 
accuracy rate improves slightly to 89.9% (231 
of 257 pixels classified correctly). The 
resultant confusion matrix can be seen in 
Figure 5. 
3.2 Classification utilizing less than six 
spectral values 
Three experiments were performed using the 155 
instance training set and different settings 
for the predictiveness threshold. In the first 
experiment the predictiveness threshold was set 
at 1.07 thereby eliminating the attribute RED 
from use during instance classification. (See 
figure 3 for predictiveness values.) When the 
remaining five attributes were used to classify 
147 instances, the classification results were 
identical to those of experiment one (see 
subheading 3.1) giving an overall accuracy of 
98.6%. 
For the second experiment the predictiveness 
threshold was set at 1.13 thereby eliminating 
the spectral attributes BLUE, RED and IR2 from 
use. The results showed a classification 
accuracy of 96.6% with 142 of 147 instances 
being classified correctly. 
In the third experiment, the predictiveness 
threshold was set at 1.17, which eliminated all 
attributes with the exception of NEAR IR from 
the classification process. The results 
obtained in using this single attribute for 
instance classification showed an overall 
classification accuracy of 55.8%. 
Three additional experiments were performed 
using the 60 instance training set and various 
settings for the predictiveness threshold. The 
predictiveness values (not shown) for the 60 
instance training set differed from those found 
in Figure 3. Specifically, BLUE was found to be 
the least predictive of class membership. NEAR 
IR and IR2 were the most predictive of class 
membership. 
For the first experiment, a predictiveness 
threshold of 1.13 eliminated the attribute BLUE 
654 
  
Figure 4: Confusion matrix, with columns 
representing actual categories of pixels and 
rows representing classifications by SX-WEB. 
  
from use during classification. Instance 
classification resulted in 21 misclas- 
sifications and gave a 91.3% accuracy level. 
In the second experiment, with a predictiveness 
threshold of 1.15, all attributes excepting 
BLUE and RED were predictive of class 
membership. The resulting classification showed 
215 of 247 instances classified correctly 
giving an 89% accuracy level. 
In the final experiment, a predictiveness 
threshold setting of 1.17 resulted in NEAR IR 
and IR2 being the only attributes predictive of 
class membership. Sixty two of the 242 
instances were misclassified giving an accuracy 
rate of 74.4%. 
3.3 Comparisons to other systems 
As a means of comparison of these results to 
those obtained from other methods utilizing 
similar data sets, the reader is directed to 
(Civco, 1992a). 
The results from (Civco, 1992a) can be 
partially summarized as follows: 
The maximum likelihood estimation resulted in 
an overall classification accuracy of 91.5%. 
A back-propagation neural network with a 6- 
element input layer, a 15-element hidden layer, 
and a l-element output layer, resulted in an 
overall classification accuracy for 468 test 
pixels of 66.7%. 
A similar network, but with both a 6-element 
hidden layer and a second 15-element hidden 
layer, resulted in an overall classification 
accuracy of 64.5%. 
It is especially interesting to note that the 
greatest number of misclassifications by SX-WEB 
(see Figure 5) were the result of misclas- 
sifying Wetland (WE) pixels as Agricultural 
(AG) pixels. This misclassification was not 
present in the results found in the neural nets 
of (Civco, 1992a), although there was evidence 
of this type of misclassification with the 
maximum likelihood technique. 
4. CONCLUSIONS AND FUTURE WORK 
Recent research (Keil, 1987; Porter, 1990) 
supports an exemplar-based approach to concept 
learning. The findings of this research lends 
additional support to an exemplar-based concept 
learning paradigm. SX-WEB's similarity measure 
and evaluation function performed exceptionally 
well in the classification of pixel images 
representing fifteen different Landsat image 
types. High classification accuracy was 
achieved even when each concept class contained 
as few as three training instances. The results 
of predictiveness testing were also positive in 
that high levels of classification accuracy 
were maintained when a limited number of 
2 sp XJ" o^ iu 
Ct m D ocr OO B S MeO Pt =" MIA 
PP TAH 9% 30 0 HH 
conmpmuurcti